Microsoft has just announced the Fara-7B, its first compact AI worker model capable of using a human-like computer through just one screenshot.
Unlike complex agency systems that rely on large cloud infrastructure, Fara-7Bs are designed to run directly on the device, helping to reduce latency, increase privacy and open up a completely new way of PC interaction.
The Fara-7B belongs to the group of small language models (SLM) that Microsoft has been chasing since last year, following the Phi series integrated on Windows 11.
However, Fara 7B is a more important step forward, built as a Computer Use Agent (CUA), the model is capable of understanding the computer interface, analyzing screenshots and performing practical actions such as clicking, entering text or web navigation.
Thanks to that, users can assign the mass-processing model popular tasks without manual intervention.
The special thing about the Fara-7B is its simplicity. Most of today's CUA models need a large cloud server fleet, multiple subsystems, and huge computing power just to analyze the screen.
Microsoft said the Fara-7B is just a single model, regardless of the complicated support model or pipeline, but still performs well on par with large-scale AI agents.
With a size of 7 billion parameters, the model can run directly on personal PC, while ensuring that users do not have to send data to the cloud.
To train Fara-7B, Microsoft built a synthetic FaraGen data system, where AI agents simulate human behavior over more than 70,000 real domain names.
Each working session includes many steps such as retesting, rolling, searching, handling errors and being assessed by three independent AI models, ensuring reasonableness.
After the filtering process, more than 145,000 sessions with more than 1 million actions were retained to train the model.
In practice, the Fara-7B consumes about 124,000 input token and 1,100 output token per mission.
The standard score of the model is also impressive, with 73.5% on Web Voyager, 34.1% on OnlineMind 2 Web, 26.2% on DeepShop and 38.4% on WebTailBench, focusing on real-life tasks such as finding a job or looking for real estate.
The Fara-7B is currently available on Microsoft Foundry and hugging Face under the MIT license. Microsoft also released an optimal quantum version for the Copilot+ PC running Windows 11, allowing the community to test directly.
With openness and localized performance, Fara-7B promises to become a platform to promote the development of AI agents that automate daily tasks.
